Elamur Research⌕
☰
home / papers / 2026 / Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection

Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection

Paper · 2026 · Yuan Zhao and Ali Abdi

Summary

LLMs are widely used as high-volume classifiers, such as hate speech moderators. Explaining their choices through generated rationales like chain-of-thought adds heavy per-item inference cost. Zhao and Abdi instead treat a black-box LLM as a rationally inattentive decision maker and infer interpretable parameters only from its input–output behaviour. Their approach builds on Sims's rational inattention (RI) theory and the discrete-choice solution of Matějka and McKay. The extended RI model adds a noise stage: character-level distortion masks hateful content with a "hiding rate" q, and the model then balances expected reward against a unit information cost λ. Using hate speech texts from Kennedy et al. (2020), the authors build eleven noise environments with random in-word edits and query four commercial LLMs zero-shot. The fitted model closely tracks observed accuracy across noise levels. λ separates the models and stays nearly fixed when the distortion mix changes, while the noise parameters absorb that change. This lets practitioners tell whether an accuracy drop comes from the input environment or from the model's own strategy, without generating extra tokens.

Key points

  • Models and data: GPT-5.2, GPT-3.5, Gemini-2.5 and Gemini-2.0 were tested as binary classifiers ("Hate Speech" vs. "Not Hate Speech"). The short social-media texts (YouTube, X, Reddit) come from the Kennedy et al. (2020) datasets.

  • Noise procedure: Each word is chosen for distortion with probability p′, from 0 to 1 in steps of 0.1, giving 11 environments. Each chosen word receives one swap, replace, delete or insert operation.

  • Model structure: The model has two states and two actions, with a latent noisy state V. Accuracy depends on q and the ratio r/λ, and the reward r is fixed at 1 so λ can be estimated.

  • Generalization test: With N = 1000 and three noise mixes (replace-dominant, swap-dominant, equal-probability), λ stayed stable within each model:

    • GPT-5.2: ≈0.345–0.399 (lowest cost)
    • GPT-3.5: ≈0.416–0.461
    • Gemini-2.5: ≈0.518–0.556
    • Gemini-2.0: ≈0.589–0.596 (highest cost)

    In contrast, the noise parameters shifted substantially: α ranged 0.295–0.594 and β ranged 2.318–3.06.

  • Limitations: Batching items caused cross-item label contamination, which the authors mitigated by grouping same-label items. The framework is restricted to binary decisions, and it assumes an equal reward across providers.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓New Jersey Institute of TechnologyNewark · United States16 works · 1,333 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksSocial Sciences · 1,954 worksSafety Research · 626 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 1,035 worksBias, fairness & safety · 587 worksInterpretability · 149 works

Cite

Cite as Yuan Zhao and Ali Abdi. 2026. Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.23

@inproceedings{zhao2026interpretability,
  title = {{Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection}},
  author = {Yuan Zhao and Ali Abdi},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.23},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.23}
}

References

13 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yevgeni Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In Proceedings of the 6th International Conference on Learning Representations, pages 1–13, Vancouver, Canada. find
  2. Filip Matějka and Alisdair McKay. 2015. Rational inattention to discrete choices: A new foundation for the multinomial logit model. American Economic Review, 105:272–298. find
  3. Koushik Pattanayak, Vikram Krishnamurthy, and Ankur Jain. 2024. Interpretable deep image classification using rationally inattentive utility maximization. IEEE Journal of Selected Topics in Signal Processing, 18:254–270. find
  4. Christopher A. Sims. 2003. Implications of rational inattention. Journal of Monetary Economics, 50:665– 690. find
  5. Bolin Wen, K. P. Subbalakshmi, and Feng Yang. 2022. Revisiting attention weights as explanations from an information theoretic perspective. arXiv preprint arXiv:2211.07714. — [arXiv:2211.07714](https://arxiv.org/abs/2211.07714) find
  6. Yujia Yang, Jihyung Kim, Yubin Kim, Nayeon Ho, James Thorne, and Se-Young Yun. 2023. HARE: Explainable hate speech detection with step-by-step reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5490– 5505, Singapore. find
  7. Andrew Caplin and Mark Dean. 2015. Revealed preference, rational inattention, and costly information acquisition. American Economic Review, 105:2183– 2203. find
  8. Andrew Caplin and Daniel Martin. 2015. A testable theory of imperfect perception. The Economic Journal, 125:184–202. find
  9. Mark Dean and Nathaniel Neligh. 2023. Experimental tests of rational inattention. Journal of Political Economy, 131:3415–3461. find
  10. Ankur Jain and Vikram Krishnamurthy. 2025. Interacting large language model agents: Bayesian social learning based interpretable models. IEEE Access, 13:25465–25504. find
  11. Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is BERT really robust? A strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8018– 8025, New York, NY.
  12. Christopher J. Kennedy, Geoffrey Bacon, Alexander Sahn, and Clara von Vacano. 2020. Constructing interval variables via faceted Rasch measurement and multitask deep learning: A hate speech application. arXiv preprint arXiv:2009.10277. — [arXiv:2009.10277](https://arxiv.org/abs/2009.10277) find
  13. Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2018. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhao-2026-interpretability-llm-classifiers-rational-inattention/